The　advantages　and　applications　of　rule-based　systems　have　caused　them　to　be　widely　recognized　as　one　of　the　most　popular　systems　in　human　decision-making,　due　to　their　accuracy　and　efficiency.　To　improve　the　performance　of　rule-based　systems,　there　are　several　issues　proposed　to　be　focused.　First,　it　is　unnecessary　to　take　the　entire　rule　base　into　consideration　during　each　decision-making　process.　Second,　there　is　no　need　to　visit　the　entire　rule　base　to　search　for　the　key　rules.　Last,　the　key　rules　for　each　decision-making　process　should　be　different.　This　paper　focuses　on　an　advanced　extended　belief　rule　base　(EBRB)　system　and　proposes　a　multi-attribute　search　framework　(MaSF)　to　reconstruct　the　relationship　between　rules　in　the　EBRB　to　form　the　MaSF-based　EBRB.　MaSFs　can　be　divided　into　k-dimensional　tree　(KDT)-based　MaSFs　and　Burkhard–Keller　(BKT)-based　MaSFs.　The　former　is　targeted　at　decision-making　problems　with　small-scale　attribute　datasets,　while　the　latter　is　for　those　with　large-scale　attribute　datasets.　Based　on　the　MaSF-based　EBRB,　the　k-neighbor　search　and　the　best　activated　rule　set　algorithms　are　further　proposed　to　find　both　the　unique　and　the　desired　rules　for　each　decision-making　process　without　visiting　the　entire　EBRB,　especially　when　handling　classification　problems　with　large　attribute　datasets.　Two　sets　of　experiments　based　on　benchmark　datasets　with　different　numbers　of　attributes　are　performed　to　analyze　the　difference　between　KDT-based　MaSFs　and　BKT-based　MaSFs,　and　to　demonstrate　how　to　use　MaSFs　to　improve　the　accuracy　and　efficiency　of　EBRB　systems.　MaSFs　and　their　corresponding　algorithms　are　also　regarded　as　a　general　optimization　framework　that　can　be　used　with　other　rule-based　systems.　©　2016　Elsevier　Inc.